Re: [PATCH] mm/memory-failure: fix VM_BUG_ON_PAGE(PagePoisoned(page)) when unpoison memory

[David Hildenbrand <david@redhat.com>]

On 18.07.24 05:04, Miaohe Lin wrote:
> On 2024/7/17 17:01, David Hildenbrand wrote:
>> On 16.07.24 04:34, Miaohe Lin wrote:
>>> On 2024/7/16 0:16, David Hildenbrand wrote:
>>>> On 15.07.24 08:23, Miaohe Lin wrote:
>>>>> On 2024/7/13 5:09, Andrew Morton wrote:
>>>>>> On Fri, 12 Jul 2024 14:42:49 +0800 Miaohe Lin <linmiaohe@huawei.com> wrote:
>>>>>>
>>>>>>> When I did memory failure tests recently, below panic occurs:
>>>>>>>
>>>>>>> page dumped because: VM_BUG_ON_PAGE(PagePoisoned(page))
>>>>>>> kernel BUG at include/linux/page-flags.h:616!
>>>>>>> Oops: invalid opcode: 0000 [#1] PREEMPT SMP NOPTI
>>>>>>> CPU: 3 PID: 720 Comm: bash Not tainted 6.10.0-rc1-00195-g148743902568 #40
>>>>>>> RIP: 0010:unpoison_memory+0x2f3/0x590
>>>>>>> RSP: 0018:ffffa57fc8787d60 EFLAGS: 00000246
>>>>>>> RAX: 0000000000000037 RBX: 0000000000000009 RCX: ffff9be25fcdc9c8
>>>>>>> RDX: 0000000000000000 RSI: 0000000000000027 RDI: ffff9be25fcdc9c0
>>>>>>> RBP: 0000000000300000 R08: ffffffffb4956f88 R09: 0000000000009ffb
>>>>>>> R10: 0000000000000284 R11: ffffffffb4926fa0 R12: ffffe6b00c000000
>>>>>>> R13: ffff9bdb453dfd00 R14: 0000000000000000 R15: fffffffffffffffe
>>>>>>> FS:  00007f08f04e4740(0000) GS:ffff9be25fcc0000(0000) knlGS:0000000000000000
>>>>>>> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>>>>>> CR2: 0000564787a30410 CR3: 000000010d4e2000 CR4: 00000000000006f0
>>>>>>> Call Trace:
>>>>>>>     <TASK>
>>>>>>>     unpoison_memory+0x2f3/0x590
>>>>>>>     simple_attr_write_xsigned.constprop.0.isra.0+0xb3/0x110
>>>>>>>     debugfs_attr_write+0x42/0x60
>>>>>>>     full_proxy_write+0x5b/0x80
>>>>>>>     vfs_write+0xd5/0x540
>>>>>>>     ksys_write+0x64/0xe0
>>>>>>>     do_syscall_64+0xb9/0x1d0
>>>>>>>     entry_SYSCALL_64_after_hwframe+0x77/0x7f
>>>>>>> RIP: 0033:0x7f08f0314887
>>>>>>> RSP: 002b:00007ffece710078 EFLAGS: 00000246 ORIG_RAX: 0000000000000001
>>>>>>> RAX: ffffffffffffffda RBX: 0000000000000009 RCX: 00007f08f0314887
>>>>>>> RDX: 0000000000000009 RSI: 0000564787a30410 RDI: 0000000000000001
>>>>>>> RBP: 0000564787a30410 R08: 000000000000fefe R09: 000000007fffffff
>>>>>>> R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000009
>>>>>>> R13: 00007f08f041b780 R14: 00007f08f0417600 R15: 00007f08f0416a00
>>>>>>>     </TASK>
>>>>>>> Modules linked in: hwpoison_inject
>>>>>>> ---[ end trace 0000000000000000 ]---
>>>>>>> RIP: 0010:unpoison_memory+0x2f3/0x590
>>>>>>> RSP: 0018:ffffa57fc8787d60 EFLAGS: 00000246
>>>>>>> RAX: 0000000000000037 RBX: 0000000000000009 RCX: ffff9be25fcdc9c8
>>>>>>> RDX: 0000000000000000 RSI: 0000000000000027 RDI: ffff9be25fcdc9c0
>>>>>>> RBP: 0000000000300000 R08: ffffffffb4956f88 R09: 0000000000009ffb
>>>>>>> R10: 0000000000000284 R11: ffffffffb4926fa0 R12: ffffe6b00c000000
>>>>>>> R13: ffff9bdb453dfd00 R14: 0000000000000000 R15: fffffffffffffffe
>>>>>>> FS:  00007f08f04e4740(0000) GS:ffff9be25fcc0000(0000) knlGS:0000000000000000
>>>>>>> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>>>>>> CR2: 0000564787a30410 CR3: 000000010d4e2000 CR4: 00000000000006f0
>>>>>>> Kernel panic - not syncing: Fatal exception
>>>>>>> Kernel Offset: 0x31c00000 from 0xffffffff81000000 (relocation range: 0xffffffff80000000-0xffffffffbfffffff)
>>>>>>> ---[ end Kernel panic - not syncing: Fatal exception ]---
>>>>>>>
>>>>>>> The root cause is that unpoison_memory() tries to check the PG_HWPoison
>>>>>>> flags of an uninitialized page. So VM_BUG_ON_PAGE(PagePoisoned(page)) is
>>>>>>> triggered.
>>>>>>
>>>>>> I'm not seeing the call path.  Is this BUG happening via
>>>>>>
>>>>>> static __always_inline void __ClearPage##uname(struct page *page)    \
>>>>>> {                                    \
>>>>>>       VM_BUG_ON_PAGE(!Page##uname(page), page);            \
>>>>>>       page->page_type |= PG_##lname;                    \
>>>>>> }
>>>>>>
>>>>>> ?
>>>>>>
>>>>>> If so, where's the callsite?
>>>>>
>>>>> It is BUG on PF_ANY():
>>>>>
>>>>> PAGEFLAG(HWPoison, hwpoison, PF_ANY)
>>>>>
>>>>> #define PF_ANY(page, enforce)    PF_POISONED_CHECK(page)
>>>>>
>>>>> #define PF_POISONED_CHECK(page) ({                    \
>>>>>       VM_BUG_ON_PGFLAGS(PagePoisoned(page), page);        \
>>>>>       page; })
>>>>>
>>>>> #define    PAGE_POISON_PATTERN    -1l
>>>>> static inline int PagePoisoned(const struct page *page)
>>>>> {
>>>>>       return READ_ONCE(page->flags) == PAGE_POISON_PATTERN;
>>>>> }
>>>>>
>>>>> The offlined pages will have page->flags set to PAGE_POISON_PATTERN while pfn is still valid:
>>>>>
>>>>> offline_pages
>>>>>      remove_pfn_range_from_zone
>>>>>        page_init_poison
>>>>>          memset(page, PAGE_POISON_PATTERN, size);
>>>>
>>>> Worth noting that this happens after __offline_isolated_pages() marked the covering sections as offline.
>>>>
>>>> Are we missing a pfn_to_online_page() check somewhere, or are we racing with offlining code that marks the section offline?
>>>
>>> I was thinking about to use pfn_to_online_page() instead of pfn_to_page() in unpoison_memory() so we can get rid of offlined pages.
>>> But there're ZONE_DEVICE pages. They're not-onlined too. And unpoison_memory() should work for them. So we can't simply use
>>> pfn_to_online_page() in that. Or am I miss something?
>>
>> Right, pfn_to_online_page() does not detect ZONE_DEVICE. That has to be handled separately if pfn_to_online_page() would fail.
>>
>> ... which is what we do in memory_failure():
>>
>> p = pfn_to_online_page(pfn);
>> if (!p) {
>>      if (pfn_valid(pfn)) {
>>          pgmap = get_dev_pagemap(pfn, NULL);
>>          put_ref_page(pfn, flags);
>>          if (pgmap) {
>>              ...
>>          }
>>      }
>>      ...
>> }
> 
> Yup, this will be a good alternative. But will it be better to simply check PagePoisoned() instead?

The memmap of offline memory sections shall not be touched, so .... 
don't touch it ;)

Especially because that PagePoisoned() check is non-sensical without 
poisoining-during-memmap-init. You would still work with memory in 
offline sections.

I think the code is even wrong in that regard: we allow for memory 
offlining to work with HWPoisoned pages, see __offline_isolated_pages(). 
Staring at unpoison_memory(), we might be putting these pages back to 
the buddy? Which is completely wrong.


... not to mention that a function called "unpoison_memory()" doing 
nothing when it finds PagePoison() is completely confusing. Last but not 
least, take a look at the number of users of PagePoison().

Likely PagePoison() warrants a cleanup, but I am not sure yet what's the 
right thing to do.

-- 
Cheers,

David / dhildenb